Abstract:

3D Gaussian Splatting (GS) has emerged as a method that achieves high-quality 3D scene representation and fast rendering, with applications in various fields. How- ever, the substantial storage requirements of complex scenes limit its practical deployment on resource- constrained platforms. In this paper, we propose a novel method, namely learnable graph-based GS compression (L-GGSC). L-GGSC introduces a parameterized graph shift operator and a systematic parameter reduction strategy to optimize the hyperparameter search space. Evaluations on three 3D GS datasets using the typical parameter of the graph shift operators demonstrate that the parameterized graph shift operator of the proposed L-GGSC has the potential to simultaneously improve data size and rendering quality against the regular graph Laplacian matrix.